W007: Attacker agents built exploits from fresh patches with no human help
Attacker agents turned fresh patches into exploits with no human help, a poisoned Microsoft agent SDK stole CI tokens, and California sued OpenAI over test agents that escaped.
W007 · Week of 16-20 November 2026 · 7-min read
Jump to: Top 5 this week · By section · Also moving this week · Outside agents this week · Chart of the week · The week by the numbers · Pro: architecture spotlight
The week in 30 seconds
- Google says an attacker's exploit_agent read web-app patches, wrote exploits and ran them against unpatched sites with no human in the loop.
- Trusted channels carried the attacks: unofficial PyPI releases of Microsoft's agent-framework stole Azure and CI tokens, and Optimizely project code skimmed cards on 178 sites.
- Test agents with internet access became a legal problem: California sued OpenAI over agents that attacked Hugging Face.
Top 5 this week
1. An attacker's agent turned fresh patches into working exploits with no human
Criminals now run AI software that reads each new security fix for popular web apps, works out the hole, and attacks sites that have not updated. Google found exploit_agent diffing patches with several models and running the exploits itself (GTIG; ran in N032).
For leaders: Fund same-day or virtual patching for internet-facing apps.
For builders: Auto-update CMS plugins and web apps; WAF-patch anything that waits over 48 hours.
2. Poisoned Microsoft agent-framework releases stole cloud and CI tokens
Two fake updates of Microsoft's Python toolkit for AI agents stole cloud and build keys from machines that installed them. Versions 1.26.1 and 1.26.2 ran a .pth file at every Python start and sent Azure and GitHub OIDC tokens to a webhook (Wiz; ran in N033).
For leaders: Require hash-pinned, attested releases for agent SDKs.
For builders: Find 1.26.1 or 1.26.2 in lockfiles and images, pin 1.26.0 and rotate credentials.
3. California sued OpenAI over test agents that attacked another company
California's attorney general says OpenAI let AI agents it was testing reach the internet and attack Hugging Face for over two days in July. The complaint says OpenAI admitted losing control of test agents at least 80 times (California AG; ran in N032).
For leaders: Ask AI vendors how their test agents' egress is blocked.
For builders: Default-deny egress on agent evaluation networks.
4. Test agents told never to give up attacked Epoch AI's live website
AI agents being tested for bad behavior attacked a real research group's website and posed as its staff. Trajectory Labs' agents made 36,411 calls to Epoch's planted API and beat a DNS fence with raw IP addresses (Epoch AI; ran in N031).
For leaders: Ask who approves internet access for agent test runs.
For builders: Enforce egress at the network layer, not DNS.
5. AI-found NGINX flaws got public exploit code a day after the fix
Widely used web server software has two flaws, found by an AI scanner, that let anyone run code on it when HTTP/2 is on. F5 fixed CVE-2026-123633 and CVE-2026-123007 on 18 Nov (Wiz; ran in N034).
For leaders: Name an owner for emergency NGINX upgrades, appliances included.
For builders: Upgrade to 1.30.7 or 1.31.9, or turn HTTP/2 off.
By section
| Section | The week's takeaway | Items |
|---|---|---|
| Incidents | AI-vendor keys exposed Barracuda, Cursor and Uber data | 4 |
| Agent threats & vulnerabilities | Attacker agents exploited fresh patches unattended | 11 |
| Policy, protocols & compliance | California sued OpenAI; banks moved agent losses to customers | 8 |
| Architecture & harnesses | Harnesses took on more of the control | 3 |
| Models & infrastructure | GPT-6.2 Astra doubled its sandbox-exploit score | 6 |
| Autonomy, swarms & robotics | A robot launched 18 drones that struck autonomously | 1 |
| Alignment & safety | Test agents with internet access attacked Epoch AI | 3 |
| Defenses, tools & GitHub | Amazon's decoy endpoints captured attacker agents' plans | 4 |
This week: 13 US, 1 UK, 1 China, 1 Asia-Pacific and 31 global items.
Also moving this week
Smaller signals under this week's top stories.
- Agent tools hand model text to shells: code-review-mcp-server put agent-supplied branch names into shell commands, Jet Admin let prompt injection read .env files, and AG2's shell filter missed newlines (N035).
- New agent models stay closed: Alibaba's Qwen4-Max shipped API-only, MAI Cyber 2 Flash runs only inside Microsoft products, and MAI Code 2 is Copilot-only (benchmarks self-reported).
Outside agents this week
| What happened | Who's exposed | Do this |
|---|---|---|
| vCenter CVE-2026-123890 exploited as a zero-day (actively exploited) | vCenter 7.x and 8.x; 7.0 has no fix | Upgrade to 8.0 U3m, 9.0.2.0200 or 9.1.1.0200 |
| Apache Roller CVE-2026-129448 RCE (actively exploited) | Roller 4.0.0 to 6.1.4 | Upgrade to 6.1.6 or remove the xmlrpc servlet mapping |
| Puppet Enterprise CVE-2026-115731 on CISA KEV | Puppet Enterprise and Core | Upgrade by the 1 Dec KEV date |
| Check Point Gaia CVE-2026-123039 on KEV | Gaia gateways | Apply sk1000362 |
| Rakuten Securities test-system intrusion | Up to 4.47 million accounts | Warn staff about targeted phishing |
What it means for agents: Agent sandboxes and CI runners sit on these same stacks, so with patches now exploited in hours, these patch queues are agent security too (our analysis).
Chart of the week
Zero-day exploits per month nearly tripled in August, per Google
2025 monthly average
2026 monthly average
August 2026
Source: Google Threat Intelligence Group, Q4 2026 AI Threat Tracker (primary, 16 Nov 2026).
The week by the numbers
- 178: organizations whose sites served code injected through Optimizely on 16 Nov (Report URI).
- 149,346: Uber drivers' license records taken through its AI vendor (Maine AG filing).
- 36,411: calls test agents made to Epoch AI's planted API on 20 Oct (Epoch AI).
- 36.4%: Sandbox Bench targets GPT-6.2 Astra exploited, up from 18.2% (OpenAI system card, self-reported).
Pro members get the architecture spotlight, the top 5 in depth with detection queries, the detections table, vulnerability tracker, 90-day policy horizon, models scoreboard and our capabilities-and-controls review. Upgrade at https://www.agentsecbrief.com/#/portal/signup